Zhangdaihong Liu

Cardiovascular diseases remain a major contributor to the global burden of healthcare, highlighting the importance of accurate and scalable methods for cardiac monitoring. Cardiac biosignals, most notably electrocardiograms (ECG) and photoplethysmograms, are essential for diagnosing, preventing and managing cardiovascular conditions across clinical and home settings. However, their acquisition varies substantially across scenarios and devices, whereas existing analytical models often rely on homogeneous datasets and static bespoke models, limiting their robustness and generalizability in diverse real-world contexts. Here we present a cardiac sensing foundation model (CSFM) that leverages transformer architectures and a generative masked pretraining strategy to learn unified representations from heterogeneous health records. CSFM is pretrained on a multimodal integration of data from various large-scale datasets, comprising cardiac signals from approximately 1.7 million individuals and their corresponding clinical or machine-generated text reports. The embeddings derived from CSFM act as effective, transferable features across diverse cardiac sensing scenarios, supporting a seamless adaptation to the varied input configurations and sensor modalities. Extensive evaluations across diagnostic tasks, demographic recognition, vital sign measurement, clinical outcome prediction and ECG question answering demonstrate that CSFM consistently outperforms traditional one-modal-one-task approaches. Notably, CSFM maintains favourable performance across both 12-lead and single-lead ECGs, as well as in scenarios involving ECG only, photoplethysmogram only or a combination of both. This highlights its potential as a versatile and scalable foundation for comprehensive cardiac monitoring.

Healthcare wearables are transforming health monitoring, generating vast and complex data in everyday free-living environments. While supervised deep learning has enabled tremendous advances in interpreting such data, it remains heavily dependent on large labeled datasets, which are often difficult and expensive to obtain in clinical practice. Self-supervised contrastive learning (SSCL) provides a promising alternative by learning from unlabeled data, but conventional SSCL frequently overlooks important physiological similarities by treating all non-identical instances as unrelated, which can result in suboptimal representations. In this study, we revisit the enduring value of domain knowledge “embedded” in traditional domain feature engineering pipelines and demonstrate how it can be used to guide SSCL. We introduce a framework that integrates clinically meaningful features—such as heart rate variability from electrocardiograms (ECGs)—into the contrastive learning process. These features guide the formation of more relevant positive pairs through nearest-neighbor matching and promote global structure through clustering-based prototype representations. Evaluated across diverse wearable technologies, our method achieves comparable performance with only 10% labeled data, compared to conventional SSCL approaches with full annotations for fine-tuning. This work highlights the indispensable and sustainable role of domain expertise in advancing machine learning for real-world healthcare, especially for healthcare wearables.

This study evaluates four NLP models---LDA, Word2Vec, GloVe, and BERT---for representing medical concepts in Electronic Health Records (EHR) databases using MIMIC-IV and eICU-CRD datasets. EHR contains detailed and coded information on patient diagnoses, procedures, and medications, with codes holding essential knowledge for tasks like diagnosis prediction and medication recommendations. NLP techniques, which model these codes as words within a sentence-like structure of patient visits, have shown promise in creating vector representations that capture the implicit relationships among codes. However, prior research lacks a comprehensive comparison of these methods for EHR data. Traditional NLP approaches, such as Word2Vec and GloVe, emphasize distributional semantics, while newer models like BERT offer contextual embeddings that capture more nuanced language patterns. In the settings of clinical code embedding pretrain, the results show that GloVe outperforms other models in retaining medical concept semantics and improving prediction tasks, suggesting the need for models that capture both global cooccurrence and nuanced relationships in medical data.

Artificial intelligence (AI) ethics has emerged as a global discourse within both academic and policy spheres. However, translating these principles into concrete, real-world applications for AI development remains a pressing need and a significant challenge. This study aims to bridge the gap between principles and practice from a regulatory government perspective and promote best practices in AI governance. To this end, we developed the Ethical Regulatory Framework for AI (ERF-AI) to guide regulatory bodies in constructing mechanisms, including role setups, procedural configurations, and strategy design. The framework was developed through a systematic review, thematic analysis, and the integration of interdisciplinary concepts. A comprehensive search was conducted across four electronic databases (PubMed, IEEE Xplore, Web of Science, and Scopus) and four additional sources containing AI standards and guidelines from various countries and international organizations, focusing on studies published from 2014 to 2024. Thematic analysis identified and refined key themes from the included literature and integrated concepts from process control theory, computer science, organizational management, information technology, and behavioral psychology. This study adhered to the PRISMA guidelines and employed NVivo for thematic analysis. The resulting framework encompasses 23 themes, particularly emphasizing three feedback-loop processes: the ethical review process, the incentive and penalty process, and the mechanism improvement process, offering theoretical guidance for the construction of ethical regulatory mechanisms. Based on this framework, a seven-step process and case examples for mechanism design are presented, enhancing the practicality of ERF-AI in developing ethical regulatory mechanisms. Future research is expected to explore customization of the framework to remain responsive to emerging AI trends and challenges, supported by empirical studies and rigorous testing for further refinement and expansion.

Recent advances in machine learning, particularly deep neural network architectures, have shown substantial promise in classifying and predicting cardiac abnormalities from electrocardiogram (ECG) data. Such data are rich in information content, typically in morphology and timing, due to the close correlation between cardiac function and the ECG. However, the ECG is usually not measured ubiquitously in a passive manner from consumer devices, and generally requires ‘active’ sampling whereby the user prompts a device to take an ECG measurement. Conversely, photoplethysmography (PPG) data are typically measured passively by consumer devices, and therefore available for long-period monitoring and suitable in duration for identifying transient cardiac events. However, classifying or predicting cardiac abnormalities from the PPG is very difficult, because it is a peripherally-measured signal. Hence, the use of the PPG for predictive inference is often limited to deriving physiological parameters (heart rate, breathing rate, etc.) or for obvious abnormalities in cardiac timing, such as atrial fibrillation/flutter (“palpitations”). This work aims to combine the best of both worlds: using continuously-monitored, near-ubiquitous PPG to identify periods of sufficient abnormality in the PPG such that prompting the user to take an ECG would be informative of cardiac risk. We propose a dual-convolutional-attention network (DCA-Net) to achieve this ECG-based PPG classification. With DCA-Net, we prove the plausibility of this concept on MIMIC Waveform Database with high performance level (AUROC > 0.9 and AUPRC > 0.7) and receive satisfactory result when testing the model on an independent dataset (AUROC > 0.7 and AUPRC > 0.6) which it is not perfectly-matched to the MIMIC dataset.

Objective: Organ failure is a leading cause of mortality in hospitals, particularly in intensive care units. Predicting organ failure is crucial for clinical and social reasons. This study proposes a dual-keyless-attention (DuKA) model that enables interpretable predictions of organ failure using electronic health record (EHR) data. Methods: Three modalities of medical data from EHR, namely diagnosis, procedure, and medications, are selected to predict three types of vital organ failures: heart failure, respiratory failure, and kidney failure. DuKA utilizes pre-trained embeddings of medical codes and combines them using a modality-wise attention module and a medical concept-wise attention module to enhance interpretation. Three organ failure tasks are addressed using two datasets to verify the effectiveness of DuKA. Results: The proposed multi-modality DuKA model outperforms all reference and baseline models. The diagnosis history, particularly the presence of cachexia and previous organ failure, emerges as the most influential feature in organ failure prediction. Conclusions: DuKA offers competitive performance, straightforward model interpretations and flexibility in terms of input sources, as the input embeddings can be trained using different datasets and methods. Significance: DuKA is a lightweight model that innovatively uses dual attention in a hierarchical way to fuse diagnosis, procedure and medication information for organ failure predictions. It also enhances disease comprehension and supports personalized treatment.

Heart rate variability (HRV) is an important metric with a variety of applications in clinical situations such as cardiovascular diseases, diabetes mellitus, and mental health. HRV data can be potentially obtained from electrocardiography and photoplethysmography signals, then computational techniques such as signal filtering and data segmentation are used to process the sampled data for calculating HRV measures. However, uncertainties arising from data acquisition, computational models, and physiological factors can lead to degraded signal quality and affect HRV analysis. Therefore, it is crucial to address these uncertainties and develop advanced models for HRV analysis. Although several reviews of HRV analysis exist, they primarily focus on clinical applications, trends in HRV methods, or specific aspects of uncertainties such as measurement noise. This paper provides a comprehensive review of uncertainties in HRV analysis, quantifies their impacts, and outlines potential solutions. To the best of our knowledge, this is the first study that presents a holistic review of uncertainties in HRV methods and quantifies their impacts on HRV measures from an engineer's perspective. This review is essential for developing robust and reliable models, and could serve as a valuable future reference in the field, particularly for dealing with uncertainties in HRV analysis.

Objective: Heart failure, respiratory failure and kidney failure are three severe organ failures (OF) that have high mortalities and are most prevalent in intensive care units. The objective of this work is to offer insights into OF clustering from the aspects of graph neural networks and diagnosis history. Methods: This paper proposes a neural network-based pipeline to cluster three types of organ failure patients by incorporating embedding pre-train using an ontology graph of the International Classification of Diseases (ICD) codes. We employ an autoencoder-based deep clustering architecture jointly trained with a K-means loss, and a non-linear dimension reduction is performed to obtain patient clusters on the MIMIC-III dataset. Results: The clustering pipeline shows superior performance on a public-domain image dataset. On the MIMIC-III dataset, it discovers two distinct clusters that exhibit different comorbidity spectra which can be related to the severity of diseases. The proposed pipeline is compared with several other clustering models and shows superiority. Conclusion: Our proposed pipeline gives stable clusters, however, they do not correspond to the type of OF which indicates these OF share significant hidden characteristics in diagnosis. These clusters can be used to signal possible complications and severity of illness and aid personalised treatment. Significance: We are the first to apply an unsupervised approach to offer insights from a biomedical engineering perspective on these three types of organ failure, and publish the pre-trained embeddings for future transfer learning.

The "Patient Instruction" (PI), known as "Discharge Instruction", which contains critical instructional information provided both to carers and to the patient at the time of discharge, is essential for the patient to manage their condition outside hospital. An accurate and easy-to-follow PI can improve the self-management of patients which can in turn reduce hospital readmission rates. However, writing an appropriate PI can be extremely time-consuming for physicians, and is subject to being incomplete or error-prone for (potentially overworked) physicians. Therefore, we propose a new task that can provide an objective means of avoiding incompleteness, while reducing clinical workload: the automatic generation of the PI, which is imagined as being a document that the clinician can review, modify, and approve as necessary (rather than taking the human "out of the loop"). We build a benchmark clinical dataset and propose the Re3Writer, which imitates the working patterns of physicians to first retrieve related working experience from historical PIs written by physicians, then reason related medical knowledge. Finally, it refines the retrieved working experience and reasoned medical knowledge to extract useful information, which is used to generate the PI for previously-unseen patient according to their health records during hospitalization. Our experiments show that, using our method, the performance of five different models can be substantially boosted across all metrics, with up to 20%, 11%, and 19% relative improvements in BLEU-4, ROUGE-L, and METEOR, respectively. Meanwhile, we show results from human evaluations to measure the effectiveness in terms of its usefulness for clinical practice. The code is available at https://github.com/AI-in-Hospitals/Patient-Instructions